Onlooker Community

Composable observability, memory, and quality-gate plugins for Claude Code — all built on the Onlooker ecosystem event substrate.

Observability substrate for Claude Code. Provides the shared $ONLOOKER_DIR storage root (default $HOME/.onlooker), canonical schema-validated event emission, session and tool tracking hooks, and prompt rules. Required by all other Onlooker plugins.

Structured session memory across context truncation: extracts decisions, dead ends, and open questions on PreCompact and reinjects the most important items at SessionStart. Requires the ecosystem plugin.

Proactive periodic auditor of the persistent instruction layer (CLAUDE.md, AGENTS.md, .claude/rules/). Discovers all instruction files in the repo, extracts semantic maps, and surfaces contradictions, shadowing, gaps, and drift before they cause expensive agent misbehavior. Requires the ecosystem plugin.

Multi-agent execution with LLM-as-a-Judge quality gates. An Actor performs work; a jury of typed Judges scores it against a project-overridable rubric; a Meta-Judge reviews the jury for bias; the gate decides accept, retry, or exhaust. Grounded in LLM-as-a-Judge (Zheng et al. 2023) and LLM-as-a-Meta-Judge (Wu et al. 2024). Requires the ecosystem plugin.

Prompt-change regression detection. When a watched agent file is modified, Echo runs a single-judge quality pass via claude -p and compares the score against a stored baseline to report improved, degraded, or neutral. Requires the ecosystem plugin.

Resource governance and budget enforcement. Tracks per-session token and cost spend, gates Task spawns before they exceed a configurable budget ceiling, and emits governor.* events for audit. Named for the steam-engine governor — a device that regulates output. Requires the ecosystem plugin.

Pre-write intent clarity gate. Intercepts write-class tool calls and requires a confidence threshold before allowing them to proceed. Evaluates the pending write against the prior assistant turn as context to avoid false positives on question-answer turns. Requires the ecosystem plugin.

Intent documentation from agent activity. Captures why changes were made — problem context, decisions, tradeoffs — and distills them into readable artifacts at session end. Requires the ecosystem plugin.

Weekly synthesis and recommendations from the full observability stack. Reads all plugin event logs, produces a structured improvement brief, and injects it at session start when the last brief is stale. Requires the ecosystem plugin.

Untrusted-content gate. Scans content flowing in through WebFetch and Read for prompt-injection patterns, and when a threat is detected closes a session-scoped gate that blocks Write, Edit, and Bash until the user explicitly clears it. Grounded in Meta's Agents Rule of Two: an agent should hold no more than two of {private data, external actions, untrusted content} at once — warden removes the external-actions property while untrusted content is in play. Requires the ecosystem plugin.

Consolidation layer between archivist's per-session artifacts and the user's durable typed memory store. Detects which session decisions, dead-ends, and open questions deserve to live across sessions, classifies them into the user/feedback/project/reference types, and queues them as proposals for explicit confirmation. Auto-promotion is opt-in. Requires the ecosystem plugin.

Maintenance layer for the user's typed auto-memory store. At every SessionStart, runs four cheap heuristic checks (date_decayed, path_broken, broken_index, orphaned_memory) against the memories at ~/.claude/projects/<encoded>/memory/ inside a wall-clock budget. Surfaces findings as a one-line pointer to /curator review; never edits the memory store directly. Requires the ecosystem plugin.

Episodic memory layer. At SessionEnd, chunks and sanitizes the session transcript and stores chunks under $ONLOOKER_DIR/historian/<project-key>/sessions/ (default $HOME/.onlooker). On UserPromptSubmit, embeds the prompt and performs similarity retrieval over stored chunks to surface relevant past context. Requires the ecosystem plugin.

Claim verification. At session end, parses the agent's final message for testable claims ("I ran the tests, they pass", "the build is green") and checks each against the actual command results in the session transcript, classifying it corroborated, contradicted, or unverifiable. Catches lying-without-malice. Advisory by default. Requires the ecosystem plugin.

Multi-session, per-project budget accounting. Rolls each session's spend into a per-project ledger on SessionEnd and surfaces "this project burned $X this week" at SessionStart. Where governor regulates a single session, bursar is the cross-session rollup: it reads governor.session.complete off the shared event bus and degrades to a session count when governor is absent. Requires the ecosystem plugin.

Per-change provenance — answers "why does this line exist?". Records session/turn metadata plus a secret-redacted, size-capped snippet for every Edit/Write/MultiEdit at PostToolUse into a per-project change ledger, then traces a file or line back to the change that introduced it by joining records to the transcripts historian preserves. Emits lineage.* events for audit. Requires the ecosystem plugin.

Per-edit lint and typecheck gate. Runs the project's configured checks on just the touched file after every Write/Edit/MultiEdit, so the agent sees its own lint and type errors before claiming success. Cheaper than running the full project verify; complements assayer (which catches the agent lying about claims). Emits inspector.* events for downstream analysis. Requires the ecosystem plugin.